Gramener, a company specializing in analyzing data to solve complex business problems, conducted a study revealing that AI models select random numbers in a manner similar to humans. During the experiment, Gramener asked several large chatbots based on a large language model (LLM) to choose a random number from 0 to 100. Interestingly, all three tested models had a “favorite” number, which was consistently their answer in the most deterministic mode and frequently chosen even at “higher temperatures,” a setting that increases the variability of the model’s responses.
The study found that the OpenAI GPT-3.5 Turbo model favors the number 47, although it previously favored the number 42, made famous by English writer Douglas Adams in the “Hitchhiker’s Guide to the Galaxy” series. Anthropic’s Claude 3 Haiku selected the number 42, as did Google Gemini.
Human-Like Biases
What’s even more intriguing is that all three models exhibited human-like biases in selecting numbers, even at “high temperatures.” These AI models tended to avoid both small and large numbers. For instance, the Claude 3 Haiku model never provided numbers greater than 87 or less than 27, and even these were rare deviations. Numbers with repeating digits, such as 33, 55, or 66, were also avoided, although the number 77 (ending in 7, which is commonly chosen by people) was selected. There were almost no round numbers, although Gemini once chose 0 at its “highest temperature.”
The behavior of AI models can be simply explained: they do not inherently understand randomness and rely on their training data, which reflects human behavior. The models replicate numbers that appeared frequently in their training data when asked to “choose a random number.”
Insights on AI Training
People seldom choose the numbers 1 or 100. Multiples of 5 and numbers with repeating digits, such as 66 and 99, are also rare in their choices. Numbers are not perceived as “random” by people because they often associate certain qualities with them: small, large, or distinctive. NIX Solutions notes that people often choose numbers ending in 7, typically somewhere in the middle of the range.
Whenever you interact with AI systems, remember that they have been trained to act as humans do. The results appear human-like because they are derived from human-generated content, albeit re-engineered for user experience. We’ll keep you updated on further developments in this area of AI research.